A NYC-based IT services firm with offices across the US came to us with a paid-social problem: their Meta acquisition had gone unstable overnight. Over the first half of 2026 we ran $847,651 of Meta spend for them, straight through the strangest paid-social environment we have ever operated in - Meta’s Andromeda shift - and their pipeline grew. This is the whole operating system we used on their account, the real numbers, and the parts that failed, written so you can copy it for your own service business.
Andromeda didn’t break Meta. It broke the old playbook. And it quietly ended the era of the growth hacker and started the era of the operator.
The 30-second version. Andromeda is Meta’s next-generation ad retrieval engine: the stage that sits before the auction and decides which ads are even eligible to compete for a person. It rewards creative diversity and clean signal; it punishes narrow audiences and cosmetic ad variations. So on this account we stopped engineering audiences and started engineering messages: distinct “creative territories,” a consolidated account, server-side conversion signal, and a tight loop between the ads and the client’s sales calls. Over six months that produced 1,463 qualified leads at $579 each, 47 new clients at roughly $18,035 to acquire, $7.8M+ in new pipeline, and a customer lifetime value of 21.64× that acquisition cost. Here is exactly how, and what you can lift for your own account.
Most advertisers met Andromeda as a feeling: results got jumpy, winners died faster, new ads either spiked or never left the gate. To fix the feeling you have to understand the machine.
Andromeda is Meta’s personalized ads retrieval engine. In Meta’s own words it is the first stage of ad selection: it narrows a pool of tens of millions of eligible ads down to a small candidate list before deeper ranking and the auction happen. It runs on NVIDIA Grace Hopper hardware, and Meta reports it delivered a +6% recall improvement and +8% ad-quality improvement on selected segments, with roughly 10,000× the model capacity and 100× the feature-extraction throughput of the systems it replaced. (Meta Engineering)
Why build a bigger retrieval engine? Because the number of ads in the system exploded. Meta says more than a million advertisers used its generative-AI tools to make 15 million ads in a single month. When Advantage+ automation and AI creative flood the auction with inventory, the bottleneck becomes retrieval: which of these millions is even worth ranking for this person, in milliseconds. Andromeda is the answer, and it reads the ad itself, the visuals, the hook, the language, the format, to decide.
Andromeda has a partner: GEM, Meta’s generative ads recommendation model, which learns patterns across organic and ad interactions and feeds ranking. Search Engine Land’s framing is the clearest we’ve seen: Andromeda decides which ads make it onto the shelf; GEM learns what shoppers buy and shapes what gets featured next. (Search Engine Land)
When we took over the account, the first symptom looked familiar. CPMs drifted up, cost per lead got inconsistent, former winners lost efficiency, and new ads either burned bright for a few days or throttled at low delivery. Every media buyer’s reflex is to call that creative fatigue and duplicate the winner.
That reflex was the trap. Fatigue was a symptom; the disease was that the account was still built for the previous version of Meta: too many manual walls, not enough distinct creative. It was telling the machine, “find more people like this, inside this narrow structure, using this small batch of ads.” Andromeda wanted the opposite: “here are many different messages, proofs, and formats. You find the match.”
Two specific things had stopped working, and they’re worth naming because most accounts still do them:
That Entity-ID trap is the part most accounts never see. When you upload ten button-color variants of the same ad, retrieval treats them as one candidate; if that candidate misses a person’s intent, all ten are excluded together. So we started auditing every new creative for visual and semantic similarity against the live set and held it below roughly 40% before it could ship. Above that line, a “new” ad just inherits an existing entity’s fatigue instead of earning its own delivery.
The question stopped being “how do we target better?” and became “how do we give Meta better options to retrieve, rank, and sequence?”
Under Andromeda, Meta infers who should see an ad largely from the ad itself. That is a profound shift for a service-business advertiser, because it means targeting now lives inside the message. You can’t just tell Meta “show this to operations leaders at mid-market companies.” You have to write an ad that only that buyer, in a specific painful moment, would stop for.
“Grow your business with our proven IT solutions.”
Generic stock visual + logo. One idea, thirty slight variations.
“Your team loses an afternoon every time something goes down. Here’s the IT partner that stops the fire drills.”
Founder-to-camera, naming the exact pain. Seven distinct beliefs, each in its own lane.
The second ad tells Meta the buyer’s pain, operating context, maturity, and likely company type, all from the creative. That’s what retrieval feeds on. This is also the moment the game tilted from hacker to operator: you no longer win with an audience trick you can copy in an afternoon. You win by understanding the buyer better than anyone else and producing that understanding, on cadence, as creative.
We didn’t make “more ads.” We rebuilt the account around a repeatable machine with seven moving parts. This is the part worth stealing, whatever service you sell.
Most accounts think 30 live ads means creative volume. Usually it’s one idea repeated 30 ways: same promise, same proof, same assumption. That barely helps Andromeda. We built around creative territories: distinct buyer beliefs, each its own lane, crossed with awareness stage, proof type, and format. For this IT services firm, the territories mapped to how their buyers felt:
| Creative territory | The buyer’s thought |
|---|---|
| IT is reactive, not strategic | “We only hear from our provider when something is already broken.” |
| Downtime keeps costing us | “Every outage burns a day of everyone’s time and we can’t predict them.” |
| Outgrown break-fix | “One internal person can’t cover us anymore.” |
| Security & compliance gaps | “A client asked for our security posture and we froze.” |
| The current MSP is a black box | “We pay every month and have no idea what we’re getting.” |
| Growth is outpacing our systems | “We’re hiring fast and IT can’t keep up.” |
| We need a partner, not a vendor | “We want someone who owns the outcome, not just tickets.” |
Every territory shipped across four layers so no two ads looked alike to retrieval: buyer pain × awareness stage × proof type × format (founder talking-head, static pain-point, carousel teardown, screen-recorded audit, short-form educational, direct-response offer). Before anything went live it passed an internal Creative Scorecard: hook strength in the first three seconds, clarity of pain and promise, proof density, ICP relevance, and offer clarity. Below the threshold, it didn’t run.
Here’s how concrete this gets. One week, a plain founder-to-camera clip where the client’s CEO read a single line from a real sales call, “we only hear from IT when it’s already on fire,” outperformed a polished, expensive edit by a wide margin. The lesson wasn’t “make uglier ads.” It was that the belief was the asset, and the belief came straight from a call, not a brainstorm.
We moved testing upstream. Instead of “Hook A vs Hook B,” we tested “which buyer belief is the most expensive and urgent right now?” A bad brief is “make five ads for the webinar.” A good brief is “make five ads for five beliefs: reactive IT, downtime cost, compliance fear, outgrown break-fix, black-box MSP.” The second brief produces signal; the first produces noise.
Fragmentation used to buy control. Under Andromeda it starves learning: too many ad sets, each with too little budget and too few creatives. We collapsed the structure into a few broad rooms and let creative do the segmenting.
| Campaign | Purpose | Structure |
|---|---|---|
| Prospecting | Find net-new demand | Broad targeting, multiple creative territories |
| Retargeting | Convert warm demand | Objection-sequenced proof by engagement stage |
| Testing | Validate new territories | Ring-fenced budget, fast read on message-market fit |
| Scaling | Push proven territories | Consolidated budget behind winners |
Budget moved on a simple 50/30/20 rule: 50% to proven winning territories, 30% to optimizing them, 20% ring-fenced for new experiments. Human judgment moved to the inputs (offer, message, territory, funnel step, qualification, measurement); Meta handled more of the distribution. Every time we tested it, consolidation beat segmentation, because pooling conversion data into fewer rooms let the system exit the learning phase faster and hold performance while we scaled spend.
For a high-consideration service, optimizing Meta for cheap leads will bleed you. A $40 lead that never books is more expensive than a $300 lead that becomes a client. So we tightened the loop between ads and reality with clean data and downstream truth:
In this era, the sales team is part of media buying. Our best creative insights came from call notes, not Ads Manager.
We retired generic “here’s a testimonial” retargeting. Each objection the client’s prospects raised got its own proof asset, sequenced by engagement stage:
We helped the client publish daily: teardown notes, founder updates, client transformations. Strong organic dwell time gives Andromeda behavioral signal, so new paid creatives skip part of the expensive cold-exploration phase. The highest-leverage version of this is simple, and we ran it relentlessly: when an organic post over-performs, promote that exact post as an ad and retarget the people who engaged.
Every territory and every ad got judged on the same downstream rule, not on how anyone felt about it that morning. Emotional attachment to a “winning” ad is what keeps a fatigued Entity ID alive past its usefulness.
To be precise: Meta didn’t ship four Andromeda “updates” this year. It shipped a string of documented changes - the Andromeda retrieval engine, the Adaptive Ranking Model, a Ranking Engineer Agent, KernelEvolve, and a Conversions-API requirement for lead optimization, among others. We group them into four shifts not by release date but by what each one actually demanded of the advertiser: better creative, faster adaptation, more volume, and cleaner downstream signal. Here is what changed at each, and what we did.
Meta’s retrieval engine reads your creative to choose candidates before the auction (Meta reports a +6% recall and +8% ad-quality improvement on selected segments), and Meta’s own guidance names creative diversification as how you give it better options. (Meta Engineering)
What we did: built distinct creative territories, a Creative Scorecard, and the sub-40% Entity-ID audit. Volume of genuinely different beliefs, not variations.
Meta documented its Adaptive Ranking Model, which routes each request to the most effective model (+3% conversions, +5% click-through for targeted users), plus a Ranking Engineer Agent that iterates ranking models autonomously. The platform now changes faster under you. (Meta Engineering)
What we did: stopped over-tuning against a moving target, consolidated, judged by downstream sales signal, and refreshed on a proactive cadence.
Meta documented 60%+ inference-throughput gains for the Andromeda Ads model, so it can serve more complex retrieval on a larger ad pool. Advertisers feel that as more competition and faster platform change. (Meta Engineering)
What we did: leaned into creative volume and kept every Entity ID distinct, so our ads earned their own retrieval paths as the pool grew.
From April 2026, Meta’s qualified-leads performance goal requires Conversions API integration. The platform wants booked and qualified signal, not cheap form-fills. (Meta Business)
What we did: we were already there. Server-side CAPI with deduplication, optimizing to booked and qualified, feeding sales outcomes back as the signal.
The 21.64× held across all four.
The win moved from a targeting trick you can copy in an afternoon to a system you have to run on cadence. The numbers below came from running it for six months, not from a clever audience.
We scaled spend month over month while holding, and improving, efficiency. Lead quality stayed high, which is the only thing that matters when the goal is signing service-company clients that stay for years.
The number that matters most is the last one. Here’s the division, since a careful owner will run it anyway. The 3.3× ROAS is in-period: realized revenue against the same six months of spend. The $7.8M is new pipeline across the 47 clients, about $165,957 each, or roughly 9.2× the ~$18,035 it costs to land one. That is year one. Because service-company clients who adopt a real IT partner tend to stay for years, full lifetime value lands near $390,280 per client, which is 21.64× that acquisition cost. Different denominators, same math. You can run your own build-vs-buy math on it. (These are one client’s real results over one six-month period. Results vary with offer, creative, budget, and industry; they aren’t typical and aren’t a promise of your outcome.)
A case study that only lists wins is marketing, not a teardown. Five things we tried on this account that failed:
If you run a service business and you’re staring at a jumpy Meta account, here is the shortest honest version of what we’d do, in order.
You don’t beat Andromeda by out-tricking the algorithm. You beat it by feeding it better inputs, and by connecting ad data to what actually closes.
We packaged the executable version of this, the Creative Scorecard with the exact thresholds we use, the creative-territory worksheet, and the objection-to-proof swipe file, into a free operator’s toolkit. Reading this tells you what to do; the toolkit hands you the files to do it with.
Here’s the uncomfortable part. Everything above is a system: creative production, scorecards, consolidation, server-side signal, a sales-feedback ritual, objection libraries, weekly discipline. And paid ads are only one input to it, sitting next to outreach, content, follow-up, and CRM. Most owners don’t lose at Meta because they picked the wrong interest. They lose because they can’t run this operating system and deliver the work at the same time. When delivery ramps, the ads and the follow-up are the first things to slip, and the pipeline dries out. The playbook above is free. The 200-plus hours a quarter it takes to run it while you’re delivering is not.
That gap is the reason Markster exists. We install and operate the whole growth engine inside your existing tools: the ads, the creative, the outreach in your voice, the follow-up, the CRM hygiene, the reporting. You approve the plan and the outputs; we run it on a pre-paid engagement, and we don’t stop until you hit the goal. It is deliberately not an autonomous “AI does your sales” gimmick. It’s an operated system, a team and AI behind it, in your voice, that you sign off on.
And this is the part agencies won’t match. An agency billing a monthly retainer or a percentage of ad spend is paid whether or not you grow; it’s structurally rewarded for staying opaque and keeping you dependent. We’re paid against your goal, we show you the numbers, and we stay on it until you’re there. That’s the operator model. The $847,651 above is us running that engine for one client, a NYC IT services firm with offices across the US. See how it would run on your business.
Andromeda is one instance of a bigger pattern: a machine now stands between you and the buyer and decides who gets shown. The same thing is happening in search. Google’s AI Overviews and conversational answers increasingly resolve a query without a click, so for a growing share of searches your buyer never scans ten blue links, they read one synthesized answer, and either your firm is named in it or it isn’t.
That makes a new question urgent for every service business: when someone asks ChatGPT, Gemini, or Perplexity for “the best [your service] near me,” are you in the answer, or is a competitor? For most firms the honest response is that they have no idea. It is the retrieval problem again, one layer up: get read and chosen by the model, or stay invisible.
It’s Meta’s AI retrieval engine that narrows a huge pool of eligible ads to a small candidate list before the auction, reading your creative to decide what’s eligible, which makes creative, not targeting, the primary lever.
No. It broke the accounts still built the old way. Tight audiences plus a handful of recycled winners went unstable; diverse creative, clean signal, and consolidation got stronger.
Less than it did. Broad targeting with specific, high-signal creative now beats hand-built interest and lookalike stacks in most accounts. The targeting effectively moved into the message.
Feed Meta downstream truth. Wire server-side conversions and optimize to booked and qualified leads instead of form-fills, then turn your sales team’s objections into your next creative.
Far less than $847,651. A few thousand dollars of disciplined testing across five to seven real creative territories will tell you which buyer beliefs are expensive and urgent. The system, not the budget, is what compounds. If you’d rather see your gaps before spending a dollar, run the free account check.